Frontier Coding Agents Can Now Implement an AlphaZero Self-Play Machine Learning Pipeline For Connect Four That Performs Comparably to an External Solver

A new arXiv paper from Joshua Sherwood, Ben Aybar, and Benjamin Kaplan benchmarks frontier AI agents on a complex, autonomous task: implementing an AlphaZero-style machine learning pipeline for Connect Four from a minimal description, running on consumer hardware within three hours. The goal is to measure AI's ability to accelerate AI research—a key safety concern. The resulting game AIs were evaluated in a round-robin tournament against the Pascal Pons external solver.

Claude Opus 4.7 emerged as the clear winner, beating the solver as first-mover in 7 of 8 trials—statistically significantly better than GPT-5.4, which won at most 2 of 8. The task, which no frontier agent could complete in January 2026, is now near-saturation. Notably, GPT-5.4 consistently used far less of its time budget, suggesting possible sandbagging; a follow-up probe with shorter prompts increased its usage, though Bradley-Terry ratings showed only directional differences.